TransferTWAS: A transfer learning framework for cross-tissue transcriptome-wide association study.

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Transcriptome-wide association studies (TWASs) are crucial for understanding the genetic basis of complex traits.
• Developing robust gene-expression imputation models for tissues with limited sample sizes presents a significant challenge in TWASs.

Purpose of the Study:

• To introduce TransferTWAS, a novel framework leveraging transfer learning to enhance gene-expression prediction in TWASs.
• To improve imputation accuracy and statistical power in multi-tissue TWAS analyses, particularly for tissues with sparse data.

Main Methods:

• TransferTWAS adaptively transfers information from multiple external tissues to a target tissue using a data-driven weighting strategy.
• The framework assigns higher weights to genetically similar tissues, outperforming methods that ignore or indirectly model tissue similarity.
• Performance is evaluated through simulations and analyses of real-world datasets (ROS/MAP, GEUVADIS) and a low-density lipoprotein cholesterol GWAS.

Main Results:

• TransferTWAS demonstrated superior imputation accuracy compared to existing multi-tissue TWAS methods in simulations.
• Analyses revealed a substantial gain in statistical power while maintaining robust control over type-I errors.
• The framework successfully identified more genetic associations for complex traits, including low-density lipoprotein cholesterol, than conventional methods.

Conclusions:

• TransferTWAS offers a powerful and adaptive approach to multi-tissue gene-expression imputation for TWASs.
• The method effectively utilizes information from genetically similar tissues, overcoming limitations of existing approaches.
• TransferTWAS enhances the discovery of genetic associations for complex traits, paving the way for more comprehensive genetic studies.